Reddit Users Report H3-Image Slims People Down While H3 Video Preserves Body Shape
poptamale · reddit · 2026-11-12
A Redditor comparing H3-Image vs the H3 video single-frame edit in ComfyUI finds both preserve faces well, but the video workflow keeps body build across pose/outfit/background changes while H3-Image slims down bigger people even with pose unchanged.
- Setup: 3090 + 64GB RAM; H3-Image pruned INT8 at 4MP, 6 steps, official reference-aware prompt expander; video edit via FL2VA at 2MP, 8 steps
- Matching the official template's anchor setup helped slightly; higher res, realism LoRAs, and prompt-expander thinking mode did not fix body shape
- OP also asks for good photo-restore workflows, currently doing a Flux.2 restore pass before editing
More from Multimodal
- Seedream 5.5 recolor test: hoodie changes color, every other detail stays put — testingcatalog · 2026-11-12
- Seedream 5.5 tested: ByteDance's image model keeps details consistent across 100 edits — testingcatalog · 2026-11-12
- Fable 5.2 studied all of music theory, then brute-forced 1000s of patterns into a 72-bar house track — deedydas · 2026-11-12
- Fable 5.2 generates 72-bar house tracks by testing thousands of patterns per instrument — deedydas · 2026-11-12
- Redditor previews first full 28-minute AI movie — ThinkingTheUnth_024 · 2026-11-12
- Anime short generated in one hour on a 4070 Super using a shared ComfyUI workflow — Pied_Film10 · 2026-11-12